Inline images

On terminals that support images, Chawan is capable of displaying various image formats inline.

Enabling images

There are actually two switches for images in the config:

buffer.images: this enables downloading images, even if they cannot be displayed. This is the switch you typically need.
display.image-mode: sets the inline image display method. Defaults to “auto”, but may also be set to “none”, “sixel”, or “kitty” manually. This switch is rarely useful.

In general, you only have to set buffer.images to true:

# in ~/.chawan/config.toml (or ~/.config/chawan/config.toml)
[buffer]
images = true

The image-mode switch is mainly useful if you want images to be fetched, but not displayed (“none” option). It may also help if “auto” fails to detect the terminal’s image display capabilities; however, I am not aware of any terminal where this can happen. (If you find one, open a ticket.)

Output formats

Supported output formats are:

The DEC Sixel format
The Kitty terminal graphics protocol

Support for other protocols (iTerm, MLTerm, etc.) is not planned. (To my knowledge, all image-capable terminals support at least one of the above two anyways.)

Support for environment-specific hacks such as w3mimgdisplay, ueberzug, etc. is not planned.

Sixel

Sixel is the most widely supported image format. See https://arewesixelyet.com to find a terminal that supports it.

Known quirks and implementation details:

XTerm needs extensive configuration for ideal sixel support. In particular, you will want to set the decGraphicsID, numColorRegisters, and maxGraphicSize attributes. See xterm(1) for details.
We assume private color registers are supported. On terminals where they aren’t (e.g. SyncTERM or hardware terminals), colors will get messed up with multiple images on screen.
Zellij advertises Sixel support, but the feature is completely broken in the current version, so Chawan specifically disables images in Zellij by default. In particular:
- Zellij itself does no Sixel detection, emitting Sixel data even on terminals that do not support it. (A fairly puzzling bug, given the developers seem to be aware of the DA1 feature.)
- On terminals that support Sixel, it fails to position images correctly, with the misplaced images completely messing up layout.
We send XTSMGRAPHICS for retrieving the number of color registers; on failure, we fall back to 256. You can override color register count using the display.sixel-colors configuration value.
Normally, Sixel encoding runs in two passes. On slow computers, you can try setting display.sixel-colors = 2, which will skip the first pass (but will also display everything in monochrome).
Transparency is supported, but looks weird because we approximate an 8-bit alpha channel with Sixel’s 1-bit alpha channel. Also, some terminals don’t emulate it correctly - when in doubt, try XTerm (which does).
Terminal scroll (both with LF/RI and SU/SD) is used extensively to avoid sending images several times. Expect troubles on terminals that do not handle this correctly. (For example, tmux is known to destroy images on scroll, so we avoid using scroll there.)

Kitty

On terminals that support it, Kitty’s protocol is preferred over Sixel. Its main benefit is that images do not have to be sent again every time a new slice of the image moves into the screen. Support for full RGBA without quantization is another benefit.

Note: at the time of writing, tmux does not support the Kitty image protocol. While it is possible to hack Kitty images onto tmux nevertheless, this requires significant concessions in display capabilities compared to the regular Kitty protocol (or even Sixel). Therefore, support for this hack is not planned.

Input formats

Currently, the supported input formats are:

BMP, PNG, JPEG, GIF (through stb_image)
WebP (through JebP)
SVG (through NanoSVG)

More formats may be added in the future, provided there exists a reasonably small implementation, preferably in the public domain. (I do not want to depend on external image decoding libraries, but something like stbi is OK to vendor.)

Codec module system

All image codec implementations are specified by the URL scheme “img-codec+name:”, where “name” is the MIME subtype. e.g. for image/png, it is “img-codec+png:”. (This indeed means that only “image” MIME types can be used.)

Like all schemes, these are defined (and overridable) in the urimethodmap file, and are implemented as local CGI programs. These programs take an encoded image on stdin, and dump the decoded RGBA data to stdout - when encoding, vice versa.

This means that it is possible for users to define image decoders for their preferred formats, or even override the built-in ones. (If you actually end up doing this for some reason, please send me a mail so I can add it to the bonus directory.)

A codec can have one of, or both, “decode” and “encode” instructions; these are set in the path name. So “img-codec+png:decode” is called for decoding PNGs, and “img-codec+png:encode” for encoding them.

Headers are used for transferring metadata (like image dimensions), both from the browser (input) and to the browser (output). Detailed description of the decoder & encoder interfaces follows.

decoding

When the path equals “decode”, a codec CGI script must take a binary stream of an encoded image on its standard input and print the equivalent binary stream of big-endian 8-bit (per component) RGBA values to stdout.

Input headers:

Cha-Image-Info-Only: 1

This tells the image decoder to only send image metadata (i.e. size). Technically, the decoder is free to fully decode the image, but this is a waste of resources; the browser will ignore any output received after headers.

Output headers:

Cha-Image-Dimensions: {width}x{height}

The size of the decoded image. e.g. for 123x456, 123 is width and 456 is height.

encoding

When the path equals “encode”, a codec CGI script must take a binary stream of big-endian 8-bit (per component) RGBA values on its standard input and print the equivalent encoded image to its standard output.

Input headers:

Cha-Image-Dimensions: {width}x{height}

Specifies the dimensions of the input RGBA image. This means that {width} * {height} * 4 == {size of data received on stdin}.

The format is the same as above; in fact, the design is such that you could directly pipe the output of decode to encode (and vice versa).

Cha-Image-Quality: {number}

The requested encoding quality, ranging from 1 to 100 inclusive (i.e. 1..100). It is up to the encoder to interpret this number.

(The stb_image JPEG encoder uses this.)

Output headers:

Currently, no output headers are defined for encoders.

Skipping copies with mmap

The naive implementation of the above system would have to copy the output at least twice when an image is resized. To skip these copies, stdin and/or stdout is (currently) a file in the tmp directory for:

decode stdin, when the image is already downloaded
decode stdout, always
encode stdin, always

This makes it possible to mmap stdin/stdout instead of streaming through them with read and write. When doing this, mind the following:

When reading, you must check your initial position in the file with lseek.
When writing, your headers are part of the output. At the very least, you must place a newline at the file’s beginning.
mmapping for write is slower than streaming the data with write. The only reason to mmap for write is to save memory when you can’t stream the output.
This is an implementation detail, and might change at any time in the future (e.g. if we add a “no cache files” mode). Always check for S_ISREG to ensure that you are actually dealing with a file. (Use io/dynstream.nim’s readLoopOrMmap and maybeMmapForSend to deal with this automatically.)